Jaehyun Cho

Good point on max_tokens. A limit tuned on one tokenizer can cut another provider's JSON reply mid-object. To answer your question: OpenAI counts are exact. The o200k_base tokenizer runs in the browser via js-tiktoken. Claude and Gemini are estimates, since their tokenizers can't run client-side. The counter takes the o200k count and applies a per-family ratio: 1.30 for Claude models since 4.7, 1.05 for Haiku 4.5 and 0.95 for Gemini. For exact Claude numbers, Anthropic's count_tokens endpoint is the reliable route. The 30% is an average for English text and shifts with language and content, so I'd still leave some headroom on max_tokens.

Thanks Ahmet, both are fair points. On the input/output ratio: agreed, the frontier moves. Models with cheap input but expensive output change places once replies get long, so a short-chat view and a long-report view would rank differently. For now the token counter on tokensave.app lets you paste your own prompt and set the expected output length, which gets close to that sensitivity check for one workload. A side-by-side ratio view is a good idea for a future update. Cost per accepted result is the better number for real deployments. A cheap model that needs two retries is not cheap anymore. The hard part is that the acceptance rate depends on the task, so I'd rather expose it as a retry factor people set themselves than publish one figure per model. Thanks for reading it this closely.